woolisbest2026/LFM2.5-1.2B-JP-202606-GGUF

LFM2.5-1.2B-JP-202606-GGUF

GGUF quantizations of [LiquidAI/LFM2.5-1.2B-JP-202606](https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP-202606).

LFM2.5-1.2B-JP-202606 is a compact Japanese-capable chat model from Liquid AI, designed for fast local inference, agentic workflows, tool use, structured outputs, and on-device applications.

This repository provides the model in GGUF format for use with llama.cpp and other GGUF-compatible applications.

Model Details

The original model was developed by Liquid AI.

Available GGUF Files

Recommended Quantization

For most users:

Q4_K_M is the recommended choice.

It provides a good balance between model quality, memory usage, and inference speed.

If memory usage is not a concern, use Q6_K or Q8_0 for higher fidelity.

If extremely low memory usage is required, Q2_K is available, although quality degradation can be noticeable.

F16 is provided for users who want the original weights in GGUF format without integer quantization.

Recommended Usage

LFM2.5-1.2B-JP-202606 is particularly suitable for:

• —Japanese conversation
• —Japanese text generation
• —English ↔ Japanese translation
• —Short-form writing
• —Summarization
• —Instruction following
• —Structured output
• —Tool / function calling
• —Lightweight agentic workflows
• —Local AI assistants
• —Edge and low-memory deployment

Its relatively small 1.17B parameter size makes it especially suitable for applications where low latency and low memory usage are important.

llama.cpp

This model can be used with llama.cpp.

CLI Example

llama-cli \
  -m LFM2.5-1.2B-JP-202606-Q4_K_M.gguf \
  -ngl 99 \
  -c 32768

Server Example

llama-server \
  -m LFM2.5-1.2B-JP-202606-Q4_K_M.gguf \
  -ngl 99 \
  -c 32768 \
  --host 0.0.0.0 \
  --port 8080

Adjust -ngl according to your hardware.

Recent versions of llama.cpp should automatically read the GGUF metadata and use the appropriate chat template.

Generation Settings

The original model documentation provides the following recommended starting settings:

temperature: 0.1
top_k: 50
repetition_penalty: 1.05

These settings are starting points and can be adjusted depending on the application.

For more creative or conversational generation, a higher temperature may be appropriate.

Context Length

The model supports a context length of up to:

32,768 tokens

For local inference, the actual usable context may depend on the inference engine and available memory.

Performance

LFM2.5-1.2B-JP-202606 is designed as a compact and efficient model.

Its 1.17B parameter size makes it suitable for:

• —CPU inference
• —Laptop inference
• —Single-GPU inference
• —Edge devices
• —Local AI assistants
• —Lightweight API servers
• —High-throughput applications

GGUF makes the model convenient to run locally with llama.cpp and other compatible inference engines.

Model Architecture

The model contains:

• —1.17B parameters
• —16 layers
• —10 double-gated LIV convolution blocks
• —6 GQA blocks
• —65,536 vocabulary size
• —32K token context length

These architectural characteristics are part of the LFM2.5 design optimized for efficient inference.

Limitations

Despite its strong efficiency for its size, this is still a 1.2B-class model.

It may struggle with:

• —Complex multi-step reasoning
• —Large-scale coding tasks
• —Long and complicated instructions
• —Knowledge-intensive questions
• —Highly precise mathematical reasoning
• —Tasks requiring extensive world knowledge

For difficult tasks, a larger model may provide substantially better results.

The original model specifies a knowledge cutoff of mid-2024.

Original Model

LiquidAI/LFM2.5-1.2B-JP-202606

https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP-202606

The original model was released by Liquid AI.

Please refer to the original model card for the authoritative model description, benchmarks, license, and usage information.

License

LFM 1.0

This repository contains GGUF conversions/quantizations of the original model and is not an official Liquid AI repository.

Please review the original model repository and license before using or redistributing the model.

Citation

If you use the underlying LFM2/LFM2.5 model in research or other published work, please cite the LFM2 Technical Report:

@article{liquidai2025lfm2,
  title={LFM2 Technical Report},
  author={Liquid AI},
  journal={arXiv preprint arXiv:2511.23404},
  year={2025}
}

Credits

• —Original model: Liquid AI
• —Model: LFM2.5-1.2B-JP-202606
• —GGUF conversion/quantization: woolisbest2026
• —Format: GGUF

Disclaimer: This is an unofficial GGUF repository. The original model and its architecture were developed by Liquid AI.